AI phone receptionist for a multi-site clinic group

A six-site clinic group was losing bookings to unanswered phones. We built an AI phone receptionist on our voice-ai and ai-automation practice that answers every call, books, reschedules and cancels appointments against the live practice-management system, and transfers to a human when the call needs one. Missed calls fell by roughly 40% and around two thirds of calls now complete without a receptionist.

Challenge

Two receptionists per site were juggling walk-ins, payments and a phone that rang most between 8 and 10 a.m. Call reports showed 25–30% of morning calls going to voicemail , and voicemail bookings rarely converted. An outsourced call centre had been tried and dropped: agents could not see the diary and patients disliked repeating themselves. The group did not want to add headcount for a two-hour peak, and it did not want a robotic IVR that made older patients hang up.

What we built

• A real-time voice agent on Twilio with streaming speech-to-text (Deepgram), a GPT-4o dialogue layer and low-latency neural speech, tuned for a calm, unhurried tone and short turns.
• Tool access to the practice-management API so the agent can check clinician availability, book, move and cancel appointments and read back confirmations, not just take messages.
• Identity checks (name plus date of birth, then a one-time SMS code for anything that changes an existing booking) and a hard rule that clinical questions are never answered, only routed.
• Warm transfer to the front desk with a written summary of the call so far, and a voicemail-to-ticket fallback out of hours.
• A call dashboard showing containment, transfer reasons and transcripts flagged for review, plus a weekly evaluation set of recorded calls re-scored after every prompt change.

Results

Across the first three months, missed calls in opening hours dropped by around 40% , and between 60% and 65% of calls completed end-to-end without staff . Median response latency sits under 1.2 seconds , which is the threshold at which callers stop talking over the agent. Reception staff now handle transfers that arrive with context instead of cold calls, and the group opened bookings on the phone until 8 p.m. without extending shifts. Related pillars: ai-chatbots for the web-chat version we added afterwards, and ai-integration for the practice-system middleware.

What made the difference

We spent the first two weeks listening to recorded calls with the practice manager and wrote the agent's scripts from real conversations, not from a template. Every irreversible action requires a verified identity, and every transfer carries a summary: the two things patients and staff said mattered most.